The HRV signal we had to improvise

Showed up Worth watching Heart-rate-variability decline (via a sleep-stress proxy)

What the guide says Falling HRV across one or several mornings signals the parasympathetic 'rest and digest' branch is depleted and a crash may be near.

How usable is it, day to day? Held up as a finding

Catches ~9 in 10 crash lead-ups, but fires on ~2 of every 3 ordinary days too. When it fires, a crash follows roughly 1 time in 37, right ~1 of 37, wrong the other ~36.

Catches: 88%; of crash lead-ups it fired on (sensitivity)
Stays quiet: 32%; of ordinary days it correctly stayed quiet (specificity)
When it fires: 2.7%; of the days it fires are followed by a crash (precision)
Lift: 1.28×; × more likely than knowing nothing

These describe how the signal behaves in daily life, framed against how rare crashes are, not whether the pattern is real. That verdict is the badge above; a usable-looking number doesn't make a signal that didn't hold real. Tier C (worth watching at best), and with only 29 crashes the margins are wide.

Where this one stands

The closest thing to an early-warning signal I have, and the one signal that survived every way of slicing the record. Modest, improvised around a watch that can't measure HRV, but it kept standing. Worth watching, not a forecast.

The shape, across the four years:

the week before a crash (pooled) an ordinary day (range)

This signal's pooled path across 26 crashes (the bold line, with its spread) against the ordinary-day range (the grey band), by day before the crash. It reads the crash, not its approach: through the week before, the crash line sits inside the ordinary-day range, and rises above it only at day 0, the crash itself. This is the one signal that held (+19.7 pp, supported). Bands are the day-to-day spread across the crashes (wide at 26); 0 is my own normal.

The week before a crash for The HRV signal we had to improvise: this signal's pooled mean vs an ordinary day's, per day, in standard deviations from my own normal (0). Bands (not shown here) are the day-to-day spread across 26 crashes.
Day	Crash week (mean)	Ordinary day (mean)
day -6	-0.2	0.3
day -5	0.7	0.5
day -4	-0.2	0.4
day -3	0.4	0.7
day -2	0.1	0.4
day -1	0.6	0.3
the crash (0)	1.2	0.4

The Forerunner 245 records no HRV, so this is tested via the night-to-night variability of sleep-window stress (Garmin derives stress from the same R-R signal). The one signal that showed up across the whole record: it cleared the conventional one-in-twenty line (+19.7 pp, p=0.029), though still below the stricter cluster-corrected line for one body at n=29. 'Showed up' is its reach, not a promotion to predictor: it catches ~9 of 10 crash lead-ups but also fires on ~2 of 3 ordinary days, so when it fires a crash follows only ~1 time in 37 (Tier C, 'worth watching'). Doubly load-bearing on the site: this is also the statistical anchor under 'the body that looks perfectly recovered'.

The numbers, for the record Single-pool SUPPORTED: +19.7 pp discrimination, p=0.029 (clears one-in-twenty; below the stricter cluster-corrected line at n=29). Sensitivity ~88%, specificity ~32%; PPV ~2.7% at the ~2% base rate, lift ~1.28x (the highest of the tested signals; still Tier C).

The curve above is the real pooled shape; the fuller write-up (how it was tested, and what it means) is coming as this finding’s page is built out. For now, the shape, the status, and the working note are the honest summary.